On Fri, 21 Aug 2020 12:25:33 +0200 Joerg Vehlow <[email protected]> wrote:
> Hi Andrew and Others (please read at least the part with @RT developers), > > > Yup, mutex_trylock() from interrupt is improper. Well dang, that's a > > bit silly. Presumably the 2006 spin_lock_mutex() wasn't taken with > > irqs-off. > > > > Ho hum, did you look at switching the kexec code back to the xchg > > approach? > > > I looked into reverting to the xchg approach, but that seems to be > not a good solution anymore, because the mutex is used in many places, > a lot with waiting locks and I guess that would require spinning now, > if we do this with bare xchg. > > Instead I thought about using a spinlock, because they are supposed > to be used in interrupt context as well, if I understand the documentation > correctly ([1]). > @RT developers > Unfortunately the rt patches seem to interpret it a bit different and > spin_trylock uses __rt_mutex_trylock again, with the same consequences as > with the current code. > > I tried raw_spinlocks, but it looks like they result in a deadlock at > least in the rt kernel. Thiy may be because of memory allocations in the > critical sections, that are not allowed if I understand it correctly. > > I have no clue how to fix it at this point. > > Jörg > > [1] https://kernel.readthedocs.io/en/sphinx-samples/kernel-locking.html There's only two places that wait on the mutex, and all other places try to get it, and if it fails, it simply exits. What I would do is introduce a kexec_busy counter, and have something like this: For the two locations that actually wait on the mutex: loop: mutex_lock(&kexec_mutex); ret = atomic_inc_return(&kexec_busy); if (ret > 1) { /* Atomic context is busy on this counter, spin */ atomic_dec(&kexec_busy); mutex_unlock(&kexec_mutex); goto loop; } [..] atomic_dec(&kexec_busy); mutex_unlock(&kexec_mutex); And then all the other places that do the trylock: cant_sleep(); ret = atomic_inc_return(&kexec_busy); if (ret > 1) { atomic_dec(&kexec_busy); return; } [..] atomic_dec(&kexec_busy); -- Steve
